The problem

From competitor's analysis and secondary research - validated with quick user interviews - the targets were clear:

• Users struggle to extract accurate information from scanned documents.
• Existing tools are too technical.
• The UI and UX needed to be easier to use and navigate, with better readability and a stress-free process.
• AI should process the documents - and help users decide afterwards.

Who it's forOperations teams in insurance, finance and logistics; data and form-entry staff; and the reviewers and admins who approve.

Ver. 1: a quick demo

Our APIs were only shown working on our own platform. A product demo needed its own space - more secure and personal, but with the same feel as the API platform. With an urgent demo coming:

• Built the keyword search demo on the platform's screen look.
• Iterated the structure in discussions and quick A/B tests: the keyword field, where the Check button lives, where Reset goes.
• More demos were coming, so functions went into tabs, not a dropdown - everything visible, one click instead of two.
• It worked for the demo, but got cramped.

Ver. 2: a real, secure product

Version 2 had to look secured and independent - an application capable of huge workflows, end to end. I studied best practices and competitors (Inscribe, Azure AI Document Intelligence, Docsumo, Hyperscience and more), then wireframed.

• Calm UI. Muted white and grey, proper spacing; colour only to signal - red to delete, blue to proceed.
• Built for many files. A table of records, clear upload states (processing, completed, retry) and multi-upload.
• One workspace. File list, document preview and results side by side - read left to right, annotations jump to the page.
• Fraud detection. Fraud and trust signals plus a heatmap of how trustworthy each part of a document is.

Tap to zoom

Deep dive: loans for teachers

An EKYC flow with two sides - the branch officer who reviews, and the teacher who applies.

• The officer gets one screen per application: what was filled, what our AI extracted, predicted and suggested - then accept, reject or escalate.
• Sign once. Signing hundreds of applications a day is tedious, so the officer uploads a signature once and the system reuses it.
• AI suggests trust: “Looks good - you may approve the loan” or “This looks like a very risky profile”; duplicates are flagged.
• The teacher sees every step upfront in a progress bar, and the screen changes in place instead of scrolling - so the long form never feels long.
• Less typing. Details are pre-filled from the uploaded documents; the applicant only checks and edits.

Tap to zoom

Ver. 3: polishing after the demos

After all the active demos with heavy deadlines, we polished the UI and UX where needed - iterating one element at a time.

• A clear risk summary at the top - “High risk! 67%”.
• Fraud signals listed one by one (editing software, name, address, date edits) next to the checks that passed.
• Accept and reject always in reach.

The result

• 85%Reduction in document fraud
• 60%Reduction in manual errors
• 40%Faster turnaround times
• End to endAutomation of KYC, cheques and onboarding

Results as per the client case studies.

The final product